Statistical software for biomedical and clinical research Survival curves, diagnostic accuracy, agreement and reference intervals — produced to the standard a reviewer expects, in a workbook a co-author can open without a licence.

Time-to-event work is the analysis most often exported to another package, and the export is where the errors start. Censoring gets recoded by hand, groups are relabelled, and a second version of the dataset stops matching the one in the manuscript.

The Medical edition draws Kaplan–Meier curves, tests the difference between groups with the log-rank test, and fits Cox proportional hazards models reporting the hazard ratio. Survival analysis is in the Medical and Ultimate editions only. Comparing survival between groups covers what the log-rank test does and does not tell you.

Reporting an AUC is straightforward. Showing that your marker beats the established one is the claim that gets challenged. Defending it needs a test that accounts for both markers being measured in the same patients. Comparing two diagnostic tests covers what the DeLong test does and when the paired form is required.

Analyse-it fits empirical ROC curves under EP24-A2 and reports the AUC with DeLong confidence intervals and a Z test. Up to ten paired or independent tests can be compared at once, under equality, equivalence or non-inferiority hypotheses. Sensitivity, specificity, predictive values, likelihood ratios, the diagnostic odds ratio and Youden’s index all come with confidence intervals. For a test with a categorical result, qualitative evaluation under EP12-A2 uses Clopper–Pearson, Wilson and Newcombe intervals rather than the normal approximation. The difference matters when the sample is small or the proportion is near one.

Correlation is still the most common wrong answer to an agreement question. Two methods can correlate almost perfectly and disagree by a clinically important amount at every concentration, because correlation measures association and not closeness. Why correlation is the wrong measure sets out the argument in full.

Analyse-it produces the difference plot with mean and median bias and their confidence intervals. Limits of agreement are drawn for constant or non-constant precision, with a linear fit where the bias depends on concentration. The allowable difference band is one you specify before looking at the data. Cohen’s kappa and weighted kappa cover observer agreement on categorical outcomes.

The method has to match the sample size. Around 120 observations per partition supports the non-parametric approach. Below that, a robust or parametric method on transformed data is usually the defensible choice, and the reason for choosing it belongs in the paper.

Analyse-it estimates intervals by parametric, non-parametric, robust bi-weight, bootstrap and Harrell–Davis methods, with confidence intervals on the limits themselves. Partition by sex, age or another factor, screen outliers with Tukey box plots, and apply log, Box–Cox or one of the other transformations where the distribution requires it. Partitioning a reference interval covers when a separate interval is justified and when it is over-fitting.

Every edition includes the full Standard edition, so the rest of the manuscript does not require a second package. In the same ribbon you have hypothesis tests, ANOVA with multiple comparison procedures, linear and logistic regression with odds ratios, contingency tables and principal component analysis.